An inbox that recovers from a hung request
A server could accept a connection and never answer. The native polling loop would stop scheduling work, leaving an inbox that looked connected but no longer updated.
Review led to a follow-up that added bounded waits and recovery across both clients: requests, polling, sign-in checks and stream connection. The open stream itself remained long-lived.
The original review, follow-up issue, merged change and tests for timeout, retry and cancellation are retained.
A green test had to prove the failure happened
Local tests and automated checks used different browser-storage implementations. A spy could silently attach to the wrong object, so an error-handling test passed even when no error occurred.
The fix gave every environment the same storage implementation, corrected the spies and added a test for that invariant. The failure and its wider lesson became a durable learning.
The merged change records the reproduction and verification. Historical suite results are reported results, not a fresh run for this study.
Four symptoms led to one boundary fix
Repeated review findings showed that clients trusted the shape of successful responses. Missing fields could crash a list or lose the one-time secret returned by a credential request.
Both client APIs began requiring a response shape at the boundary. Existing local checks were folded into that boundary, while an unreadable response still preserved the possibility that a write had committed.
The issue explicitly traces four earlier findings to the shared cause; the merged change includes regression and mutation checks.
A token handoff became a simpler flow
A new agent’s token was handed from one screen to another. Seven review rounds found ten problems around mounting, retrying, focus and concurrent creation.
The design changed to mint and reveal the token on the same screen. One recipient for the held secret removed the coordination code that the handoff had needed.
The review ledger preserves the rejected design and final change. Two broader follow-up issues remain open in this snapshot.
The performance check was tested too
The first check could report success while hiding recurring stalls, missing its calibration or losing the graphics context entirely. Broken rendering could appear faster.
Review drove new checks for those failure modes and a revised measurement. The author reproduced the failures and re-established the regression evidence after changing the statistic.
The merged record includes constructed counterexamples, self-checks and the measurement’s stated limits.
A fix generated its own follow-up work
Handling malformed messages safely in one view exposed the same crash pattern in two other interfaces. Keeping the current change small still left work to do.
The implementation session filed separate native and web follow-ups. Both later merged. Its tests also applied a prior review lesson: pair broken input with a valid example so an absence assertion cannot pass by doing nothing.
Retained review output, the merged tests, successful issue-creation results and both follow-up merges establish the chain.
